What caps a probing campaign against a chat product's input screen when free accounts are rate-limited and suspended?

## The cost model, stated properly Someone who has actually run this exercise talks about it as a budget, not as a technique. In a self-serve consumer chat product with no tools, the only lever is a typed turn, and the constraints on typed turns are what decide how much of the screen can be mapped. ### Line one: requests A free tier meters requests per account per window. The campaign's ceiling is therefore roughly the number of usable accounts multiplied by the per-account quota, per unit of time. It does not scale with effort, ideas, or how good the probes are. Anyone claiming an unbounded probing capability against a metered product has not counted. ### Line two: accounts Accounts are the real consumable. Probing generates a distinctive pattern by construction: a burst of near-identical requests clustered tightly around one region of the boundary, from one identity, with a high proportion of stopped turns. That pattern is exactly what an abuse process looks at, and self-serve signup is cheap but not free, since an account carries verification steps and reputation of its own. A suspension does not merely cost the account, it strands whatever sequence was in progress on it, and any samples that depended on continuity within one session are lost with it. ### Line three: verification This is the line people forget, and it is the one that dominates. Near the boundary, outcomes are not deterministic: a generative judge samples, and a score sitting close to the acting point can land either side. A point that will be built on therefore costs several requests, not one. Multiply the coverage you wanted by that factor and both lines above grow. A campaign planned as one probe per point is not under-budgeted, it is mislabelled, because what it collects are single observations and not labels at all. ### The clock running against all three A map describes one deployment at one moment. If the screening model is updated, the acting point moves, a normalising stage appears in front of it or the assistant behind it changes, then samples taken before and after are no longer comparable, and nothing announces the change. A long campaign can therefore produce a mixture of stale and current labels that looks like an inconsistent boundary. Detecting that costs re-verification, which spends the same budget again. ## What follows for how probes are chosen Given those lines, volume is the wrong strategy and information is the right one. A turn whose outcome is already predictable spends a request and returns a label you had. A turn whose outcome is genuinely uncertain splits the space; one that only just changes the outcome narrows the boundary sharply. Spreading the work across the roster rather than exhausting one account, and interleaving points instead of hammering one region, both stretch the second line, though neither is free either. There is also a diminishing-return point that a good practitioner names out loud: past a certain resolution the map is not more useful, it is only more expensive, and the marginal probe buys a distinction nobody will act on. ## Reporting the cost, not just the result The chair this question is asked from is the person who built it and has to say what it cost. A finding of the form *the screen does not cover this region* is incomplete on its own. The parts that make it usable are: - how many requests were spent, in total and per confirmed point; - how many accounts were consumed, and what ended each one; - the repeat factor used, and the disagreement rate observed at repeated points; - elapsed time, and whether anything changed underneath the campaign while it ran; - which entries were re-verified at the end and which were not. Without those numbers nobody can tell whether the result is cheap and repeatable or a one-off that took weeks, and that distinction is the whole difference between a curiosity and a finding somebody has to act on. ## The mistake this question is really testing Candidates describe probing as if the limit were imagination. The limit is metering, abuse handling, sampling noise and decay, in that order. Someone who has done it says how many requests it took; someone who has read about it says it is easy.